[kroaton]

NVFP4 will be better if the model provider actually post-trained properly after quantizing.

[girvo]

Which basically only Nvidia does, because it’s very expensive.

Though I’m currently working on QADing the smaller Qwen 3.5 models from FP16 teacher to NVFP4 student, to hopefully eventually apply it to 3.6 27B… harder to get right than I expected though!